Yuncong Yang

M.Sc. student, Underwater Engineering Institute

School of Ocean and Civil Engineering, Shanghai Jiao Tong University

Underwater salvage operations: a ROV detecting, reaching, grasping and salvaging across three deep-sea target scenarios
The setting of the work: near-field heavy-load salvage, where the optical attenuation of the water and the dynamic interaction between vehicle and object are the two things that dominate the problem.

About

I am a master's student at Shanghai Jiao Tong University, advised by Dr. Xuyang Wang. My work is on world models for robot manipulation in the underwater setting — building predictors that reason about how a salvage ROV and its grippers interact with an object when vision is poor, contact is unobserved, and the vehicle's dynamics lag behind its commands.

Before that I studied Naval Architecture and Ocean Engineering at Dalian University of Technology, where I worked on bionic amphibious vehicles and linkage-driven grippers.

C3-JEPA architecture: DINOv3 encoder into object-centric slots, a control-conditioned latent predictor, and a behaviour agent trained on imagined rollouts with critic-based scoring
How the model is put together: multi-view frames are encoded into object-centric slots (DINOv3 + slot attention with SIGReg), a light predictor rolls the slots forward under control conditioning, and a critic scores imagined trajectories to train the behaviour agent.

Research interests

Publications

Underwater C³-JEPA: An Object-Centric Cross-View World Model for ROV Salvage
Y. Yang, et al. — under review, IEEE Transactions on Automation Science and Engineering (T-ASE)Under review
arXiv:2609.30214  ·  code
Rapid Salvage of Deep-Sea Underactuated Manipulators: Real-Time Solution of Contact Forces and Impact Assessment of Disturbance Forces
Y. Yang, et al. — Robot (《机器人》), 2025. In Chinese; EI-indexedAccepted
Multibody Dynamics and Residual Correction for Underactuated Linkage Grippers
Y. Yang, et al.In preparation

Research experience

World-model-based planning for deep-water salvage ROV (C³-JEPA)
Master's thesis project · National Key R&D Program of China (2023YFC2809701) · 2024 – present
  • Benchmarked competing world-model designs — Dreamer-style latent imagination, DIAMOND diffusion models and pure JEPA — before converging on C³-JEPA, a cross-view, control-conditioned and context-extended design.
  • Built a ROS–Gazebo data foundation (experimental pool, domain randomisation, teleoperation GUI) and collected a large synchronised multi-modal dataset across multi-camera vision, proprioception and all 12 thruster / 6 gripper control channels.
  • Designed the observation module (frozen DINOv3-L features, temporal slot attention, SIGReg) and a lightweight object-centric predictor that fuses several camera views under control conditioning.
  • Built a flow-matching diffusion planner with spherical-coordinate encoding and delay-compensated label shifting, plus a block-causal transformer rollout with critic-based trajectory selection.
  • Learned slots isolate the task object and the grippers where unguided self-supervision fails outright; C³-JEPA with the flow-matching planner cuts terminal positioning error by roughly a third and ranks candidate trajectories well above chance.
Grey-box inverse dynamics for the salvage ROV
Branch study derived from the world-model project · 2024 – present
  • Formulated a physics-informed inverse-dynamics module coupling Fossen's 6-DOF equations with learned hydrodynamic parameters and a gated residual network, recovering thruster commands from observed vehicle motion.
Underactuated salvage gripper: kinetostatics, multibody dynamics and residual correction
2024 – present
  • Built a virtual-work-based kinetostatic model of a double four-bar underactuated gripper giving fast closed-form contact-force and load-capacity solutions, and used it to assess environmental disturbance forces on the grasp.
  • Validated the solver against ADAMS simulation and an experimental gripper prototype; derived the closed-form multibody dynamics via screw theory, resolving closed kinematic loops with non-minimal coordinates and Lagrange multipliers.
  • Coupled the physical model with a parameter network and gated residual correction to absorb contact, friction and unmodelled hydrodynamic effects.
DLROV UVMS system structure: surface support system, underwater vehicle system with propulsion, perception and navigation arrays, and a manipulator system with the gripper array
The vehicle the model runs on: a work-class ROV with a propulsion array, a perception array, a navigator, and a hydraulically driven gripper array — the manipulator side is what my mechanism work models directly.

Project video

A four-stage walkthrough of the project: simulation and theory, data collection at scale, world-model prediction under test and validation, and the behaviour agent operating under the model. (1 min 38 s, with sound.)

Patents

Education

Shanghai Jiao Tong University
M.Sc. in Civil and Hydraulic Engineering · 2024.09 – 2027.06 (expected)
Dalian University of Technology
B.Eng. in Naval Architecture and Ocean Engineering · 2020.09 – 2024.06 · ranking 14/89

Contact

Underwater Engineering Institute, Shanghai Jiao Tong University
800 Dongchuan Road, Minhang District, Shanghai 200240, China

aim404@sjtu.edu.cn  ·  github.com/Aim404/c3-jepa